Author: Ben Thompson
Translation: Deep Tide TechFlow
Deep Tide Insight: Anthropic's new model, Fable, was urgently halted by the U.S. government just two months after its release. On the surface, it's about "security leaks," but in reality, it exposes a dual war between AI labs, the government, and the software industry. This company, which sells itself on "safety," is turning the safety narrative into a commercial moat. What they are really after is the user data currently held by companies like Microsoft.
I understand the cynics' perspective. They always think Anthropic's public statements—especially those accompanying model releases—are marketing-fueled fearmongering. Two months ago, Anthropic announced the launch of Mythos Preview, claiming the model was too dangerous to release publicly, particularly due to its powerful cybersecurity capabilities. Then, two months later, the company publicly released Fable, a version of Mythos with various safety guardrails added.
Based on my limited experience using it, Fable is indeed an excellent model. It's becoming difficult to objectively assess models beyond programming performance, but subjective feelings remain. I found interacting with Fable to be an outstanding experience; it made other models, including GPT 5.5 and Opus 4.8, seem small and dumb in comparison. I've only had this feeling twice before: once with GPT-4 and once with Grok 4—both represented a new generation in terms of foundational model scale and complexity. I believe Fable originates from new pre-training and is the first of a new generation.
Therefore, I fully accept that Fable/Mythos might indeed be much better at identifying and exploiting security issues, justifying Anthropic's cautious rollout. But the problem with publicly releasing a model is that guardrails can be bypassed, and apparently, this happened not long after the release.
Anthropic Confronts the U.S. Government Again
What happened next is somewhat unclear. Anthropic wrote in a blog post:
The U.S. government invoked national security authority, issuing an export control order suspending access to Fable 5 and Mythos 5 for all foreign nationals, both within and outside the United States, including Anthropic's foreign employees. The practical effect of this order is that we had to abruptly disable Fable 5 and Mythos 5 for all customers to ensure compliance. Access to all other Anthropic models remains unaffected.
We received the government's directive today at 5:21 PM ET. The letter did not provide specific details of the national security concerns. We understand the government believes a method to bypass or "jailbreak" Fable 5 has been discovered. We reviewed demos that used this specific technique to identify a handful of known minor vulnerabilities. These vulnerabilities all appeared relatively simple, and we found that other publicly available models could also discover them without requiring a bypass.
Anthropic went on to argue that non-general jailbreaks are inevitable and limited in scope, with no evidence of a general jailbreak; the discovered jailbreak appears to have been reported by Amazon, which is notable because Amazon is both an investor in Anthropic and a primary provider of the company's inference services. As I write this, Anthropic executives are in Washington D.C., trying to resolve what they insist is a misunderstanding but what White House officials hint is company leadership's indifference to legitimate national security concerns.
Given the many contested facts, I don't have much to add about the current conflict; but I'm not surprised it's happening. As I explained in "Anthropic and Alignment," conflict between the U.S. government and Anthropic was inevitable. For that matter, those who think Mythos isn't powerful enough yet to warrant such drastic government action are missing the point: if it's not powerful enough now, the next one will be, or the one after that, especially now that models are becoming increasingly useful at creating their successors.
However, this leads to another question—one that seems to validate the cynics' view: If Mythos is so dangerous, why release Fable in the first place? Why fight the government on doing what you claim to want? In fact, I find Anthropic's behavior perfectly understandable; what's unique about the company is how it justifies these actions, and it's precisely these justifications that give cynics fuel and give Anthropic its magic.
Economic Inevitability
In the early years of AI, the most economic value flowed to compute power, for obvious reasons: we didn't have enough supply to meet demand, which meant prices soared; the biggest beneficiaries were NVIDIA, TSMC, and memory makers (SK Hynix, Samsung, and Micron). Meanwhile, Anthropic and OpenAI collectively lost tens of billions of dollars building frontier models, which, once released, were distilled and commodified by open-source models, mostly from China.
This represents the pessimistic scenario for the labs—they can never cover their costs because their differentiation is fleeting, and free alternatives become "good enough"—which I believe is plausible. In a world of interchangeable models, models are commodities, and most of the value flows elsewhere. Right now it's compute, but over time, when we have enough compute, the most valuable place in the value chain will be where it has always been: owning the user touchpoint.
Therefore, there is an economic inevitability for frontier labs to get closer to users, which has always been clear to me. If you own the user touchpoint, then you have meaningful lock-in, and the best way to own the user touchpoint is to become the canvas for everything they need to do. This, in turn, means frontier labs are heading for a collision with software companies: it's the software that owns the user touchpoint, and the frontier labs' long-term interest is not simply to be a commodity input for software, but to directly replace it.
Meanwhile, software companies are striving to do the opposite. Satya Nadella outlined his vision for how companies should build on models in a post on X:
Every company must build what I call human capital and token capital. Human capital includes its employees' knowledge, judgment, relationships, ingenuity, and pattern recognition, while token capital is the AI capabilities a company builds and owns. Importantly, as token capital grows, human capital does not become less valuable. It only becomes more valuable! I believe human initiative will be the driver of token capital growth. Humans will set ambitious goals, connect dots across domains, build relationships, and identify the most important patterns. Without human guidance, your compute is idling.
This means the real opportunity isn't in choosing the best model, but in building learning loops on top of models that allow human and token capital to compound. You can outsource a task, even a job, but you can never outsource your learning. The future of a company is enabling that learning to compound between people and AI. This requires a new architectural approach that allows every business to build agent systems that improve over time while still retaining control over their intellectual property. Companies should be able to swap out 'general' models without losing the 'company veteran' expertise built into their learning systems. This is a key 'test' for your control and sovereignty in the age to come.
Nadella prefaced this vision with a warning:
What none of us want to see is a world where every company in every industry cedes value to a handful of all-consuming models. If all value is captured by just a few models, the political economy simply won't tolerate it. Society will not grant license for an AI future that hollows out entire industries.
Think about what happened in the first stage of globalization, where entire industrial economies were hollowed out by outsourcing. On the surface, GDP numbers looked good, but the displacement was real, and the consequences are still felt today. Let's not bring that dynamic into the AI era, where a handful of AI systems capture all the economic returns while entire industries find their knowledge commoditized right under their noses.
The problem with this analogy is: Globalization did happen, and industrial economies were hollowed out. It's possible this isn't a warning but a prophecy; no wonder Nadella is sounding the alarm, as Microsoft could be one of the victims. Similarly, the economic inevitability for model makers is precisely to achieve this.
Data Inevitability
These models—even Mythos—are not there yet. What they need, besides more compute, is more and better data. Model improvements increasingly come from reinforcement learning; some of that can be generated synthetically, but the most powerful lever for frontier labs is real-world use.
I think this is a primary reason both OpenAI and Anthropic offer heavily subsidized subscription plans. SemiAnalysis recently estimated that the $200 plan gets you $8,000 worth of Claude tokens and $14,000 worth of Codex tokens. Of course, both are competing for user and developer mindshare, but they are also competing for access to real usage data to improve their models.
Anthropic upped the ante significantly with Fable, announcing they will retain all data used for 30 days, even for enterprise plans that previously promised zero data retention. The company says they won't use this data for training, but they haven't put any safeguards in place to guarantee they won't in the future (like storing data with a third party). If this policy change (when Fable is restored) doesn't lead to significant customer churn, I suspect it's only a matter of time before they start using the data: it's too valuable for their ultimate goal.
Also note the virtuous cycle with moving up to the user touchpoint: the more workflows completed directly with Claude or Codex, the more data each company gets that can be fed back into training, making their product more powerful and useful, expanding the number of workflows they can serve, and expanding their access to data.
Nadella emphasizes the importance of this data in his piece, but naturally believes it should be independent of the models:
Companies need to convert workflows, domain knowledge, and accumulated judgment into AI systems that improve with every use. Private evaluation should capture whether models are truly improving on outcomes important to the business (not just external benchmarks!). Private reinforcement learning environments should make models stronger on real trajectories within the organization. Its knowledge base makes institutional memory queryable and token use more efficient.
This loop becomes the company's new intellectual property. I see it as a hill-climbing machine. Unlike most assets, it compounds. Each improved workflow generates better training signals, accelerating the accumulation of tacit knowledge unique to the company. Companies that build this early will have advantages that are difficult to replicate, regardless of any new individual model capabilities.
However, what if companies submitting to Anthropic's data policies get better results right now? Or if existing companies resist, leaving an opening for new companies—or the model makers themselves—to beat them in the market? Anthropic is certainly testing the resolve Nadella calls for.
A Claim to Power
Astonishingly, the data retention policy around Fable/Mythos wasn't even the most controversial part of the release. Instead, Anthropic stated at launch that Fable's performance would be quietly degraded if it was used for LLM development; the system card read:
We also added protective measures related to frontier LLM development. As discussed in Section 6.1 of our February 2026 Risk Report, we are concerned about risks from accelerating the overall pace of AI development, though we remain uncertain about the severity of these risks. In particular, our concern lies—as we wrote at the time—"in accelerating the ability of other AI developers to build powerful AI systems with risks similar to ours—without necessarily having corresponding protective measures."
Given recent models' ability to accelerate their own development, we have implemented new interventions limiting Claude's effectiveness on requests targeting frontier LLM development (e.g., building pre-training pipelines, distributed training infrastructure, or ML accelerator design). Using Claude to develop competing models already violates our Terms of Service, but enforcing this restriction through protective measures avoids accelerating those actors most willing to violate those terms.
Unlike our interventions for cybersecurity, biochemistry, and distillation attempts, these protective measures are invisible to the user. Fable 5 will not fall back to another model. Instead, the protective measures will limit effectiveness through methods like prompt modification, steering vectors, or Parameter-Efficient Fine-Tuning (PEFT). These interventions will not affect the vast majority of programming work. We estimate they will affect approximately 0.03% of traffic, concentrated in less than 0.1% of organizations. When these interventions are active, we expect their impact on model behavior to be minimal beyond limiting its effectiveness for developing frontier LLMs. Claude will still respond helpfully to user requests. We will continue to improve the precision of our detection methods after this model's release.
Anthropic walked back this change—Fable will now offload LLM-related requests to Opus 4.8 and disclose this offload to users—but I find the original policy highly revealing. On one hand, I don't really blame Anthropic for not wanting to help competitors; on the other hand, it should be very clear that Anthropic believes no one but them should be making frontier LLMs.
What makes this policy even more striking is that it was enacted just two months after Anthropic's dispute with the War Department: the latter wanted to use Claude for any lawful purpose, while the former wanted stricter controls on surveillance and autonomous weapons. This degradation measure represents both Anthropic's ability and willingness to quietly alter its model to enforce its policy preferences. In other words, Anthropic actively validated some critics' biggest concerns about it as a supply chain risk.
However, the broader takeaway from that episode is that Anthropic believes they should have the final say over how Anthropic is used; given they believe only they should develop frontier AI, then they effectively believe only they should have the final say over AI overall. When you combine this realization with the company's statements about AI being capable of all economic activity, you realize that Anthropic's leadership essentially wants power over everything and everyone.
The Safety Narrative
Of course, Anthropic would never phrase it so bluntly; instead, the story is about safety:
I expect Anthropic will increasingly expose its model capabilities to end-users through endpoints increasingly tailored to different workflows, even as they begin restricting the API. This substitution for software and restriction of access will be done in the name of safety, even as Anthropic fulfills its economic imperative to get closer to the end-user.
Anthropic's explanation for its significant data retention policy change is safety. Specifically, the company claims that retaining all user data for 30 days is necessary to prevent the jailbreaks the U.S. government fears. I can certainly imagine a future where safety factors also compel them to train on this data to better defend against malicious use.
Anthropic's entire origin story is rooted in the founders' belief that OpenAI wasn't taking safety seriously enough; the company believes only they can be trusted to control AI, and because they uniquely care about safety, they are justified in trying to control everyone else, including the U.S. government.
The thing about these safety justifications is this: I think they work because, for Anthropic, they are not justifications. The company genuinely believes they are the only ones who believe in superintelligence and thus are the only ones sufficiently focused on the dangers. This excuses decision after decision, policy after policy, confrontation after confrontation that, to outsiders, seem like a strange mix of cynicism and naivety.
The contrast with OpenAI is stark: One way to understand how and why OpenAI lost its lead is that, in the years following ChatGPT's release, the company was at war with itself internally, a former research lab suddenly burdened with becoming an accidental consumer tech company; as OpenAI resolved this conflict, it bled enormous talent to companies like Anthropic.
Anthropic, on the other hand, has perfect alignment between talent, mission, and business. The company can sell researchers the vision of creating a machine god, with the aura of being the kind of people who care about the dangers and are smart enough to navigate them on behalf of humanity; and every resulting policy change happens to be good for business, which is the most wonderful coincidence in the world.
I both respect and fear this alignment. I respect it because it's clearly very effective; the closest analogy might be Apple, a company that always wraps every self-serving action in the guise of doing the right thing for the user—and often they do. So does Anthropic. However, I fear that letting people convinced they know best build a smartphone I can accept or reject is one thing; letting them build superintelligence with the potential to rival or surpass the power of nation-states, or simply large corporations, is far more concerning. The history of clever people convinced they know what humanity needs is sordid, precisely because they convinced themselves the intentions were good, providing a rationale for actions that weren't.
